Classifying marine mammals signal using cubic splines interpolation combining with triple loss variational...

Area of Science:

• Acoustic Signal Processing
• Machine Learning for Marine Biology
• Bioacoustics

Background:

• Traditional time-frequency methods (Mel, STFT, WT) for passive sonar have limitations in resolution and information loss for long-term data.
• Extracting characteristic frequencies from acoustic signals, especially marine mammal communications, requires advanced signal processing techniques.
• Existing Auto-Encoder (AE) models face challenges with discontinuity and completeness during data decoding.

Purpose of the Study:

• To develop an improved method for classifying marine mammal (MM) communication signals.
• To enhance the resolution and reduce information loss in acoustic signal analysis.
• To overcome limitations of traditional time-frequency transformations and existing deep learning models.

Main Methods:

• A two-stage approach combining Cubic-splines interpolation (CSI) pre-processing with a Siamese Neural Network-Variational Auto Encoder (SNN-VAE) model.
• Generation of STFT-CSI spectrograms to strengthen relationships between characteristic frequencies.
• Stacking Mel, STFT-CSI, and Wavelet spectrograms into a feature spectrogram for comprehensive analysis.

Main Results:

• The proposed STFT-CSI spectrograms enhance frequency-based feature connectivity.
• The SNN-VAE model achieved improved classification accuracy for marine mammal signals.
• Classification accuracy increased by 11% and 20% compared to AE (2013), and by 6% compared to Resnet (2022) on the NOAA dataset.

Conclusions:

• The combination of CSI pre-processing and SNN-VAE offers a superior method for marine mammal signal classification.
• The developed technique effectively addresses resolution and information loss issues in long-term acoustic data.
• This approach provides a more robust and accurate solution for analyzing underwater acoustic communications.